Cyber-physical method and system for mechanical equipment health prediction

Abstract

The application discloses a kind of information physical fusion method and system of mechanical equipment health prediction, wherein the method comprises: collecting the working state data of mechanical equipment;Working state data is input to pre-trained information physical fusion system CPS hybrid model, and the health prediction result of mechanical equipment is obtained, wherein CPS hybrid model is obtained according to the fusion of signal processing physical modeling and deep learning data modeling;The health prediction result is mined, and the real-time state of mechanical equipment is evaluated according to the mining result, and the state evaluation result of machine equipment is obtained.The application aims to use the optimized physical modeling method to extract more rich prior knowledge, and then realize the improvement of hybrid modeling accuracy through the optimized data modeling method, which is beneficial to support predictive maintenance activities, so as to realize the monitoring and control of CPS from physical components to network components to physical components of mechanical equipment in network collaborative manufacturing and intelligent factory.

Description

Technical Field

[0001] This invention relates to the field of planning and mechanical equipment health prediction technology, and in particular to a cyber-physical fusion method and system for predicting the health of mechanical equipment. Background Technology

[0002] Cyber-Physical Systems (CPS) are an indispensable part of realizing smart manufacturing and Industry 4.0, and are gradually changing the landscape of global manufacturing. CPS is a multidisciplinary system that combines computing, communication, and control technologies to perform real-time measurement, data transmission, monitoring, decision-making, and feedback control on widely distributed embedded computing systems. The deployment of CPS in networked collaborative manufacturing and smart factory technologies has attracted considerable interest from scholars and researchers. Health prediction (including condition monitoring and fault diagnosis) of key mechanical equipment in networked collaborative manufacturing and smart factories is beneficial for ensuring the safe operation of workshops, enterprises, and supply chains, playing a crucial role in stable production. Due to the ubiquitous sensing and rapid computing capabilities of CPS in mechanical equipment within networked collaborative manufacturing and smart factories, and the increasing number of users of the Industrial Internet, data utilization has surged, leading to the rise of big data in mechanical equipment.

[0003] The main methods of cyber-physical modeling include physics-based modeling, data-based modeling, and hybrid modeling that combines both. In recent years, researchers have made some progress in applying these modeling methods to the health prediction of mechanical equipment, but problems still exist, such as the difficulty of physical modeling in meeting the needs of complex systems, and the low accuracy of data modeling and hybrid modeling.

[0004] There are two main types of CPS modeling methods: physics-based modeling and data-based modeling. Physics-based modeling uses underlying physical relationships to derive their mathematical relationships; the main advantages of physics models are their interpretability and scalability. Interpretability means that the physics model can be better understood from input to output; it is a white-box model where the causal relationships between parameters and variables are clear. Scalability is reflected in complex CPS systems composed of many subsystems. This approach can be used to model any system from simple to moderately complex. Data-based modeling eliminates the limitations of relying on system dynamics knowledge, relying entirely on data and applicable to modeling different types of systems. It is noted that while deep learning has had a significant impact on fields such as natural language processing, computer vision, and speech recognition, its impact on CPS has only recently been understood. To realize the potential of deep learning frameworks in CPS applications, it is a growing consensus that data-based machine learning modeling methods need to be combined with physics-based modeling techniques. A physics- and data-driven hybrid CPS modeling method for health prediction of mechanical transmission components is proposed, and the effectiveness of the hybrid modeling is verified experimentally.

[0005] Research has revealed that for physical modeling, considering the complexity, uncertainty, and time-varying characteristics of the operating conditions of mechanical equipment in networked collaborative manufacturing and smart factories, the physical model is often simplified to a coarse model, which is an incomplete representation of the physical processes of the real system. This is a time-consuming, expensive approach that requires expertise in the industrial application domain. For data modeling, mechanical equipment in networked collaborative manufacturing and smart factories is a complex system, and its big data often contains a large amount of redundant information. Without physical preprocessing, it is difficult to extract valuable state features. Using appropriate data preprocessing can improve prediction accuracy. However, data-based modeling does not consider the causal relationships of system variables, and they lack interpretability. Therefore, misusing data and deep learning models may lead to large errors, or even results that violate physical laws. For hybrid modeling, inserting physical preprocessing into data-based deep learning is highly feasible, but it suffers from low accuracy. Therefore, how to fully utilize physical preprocessing and data-based deep learning to further optimize the accuracy of CPS hybrid modeling has become a key focus for researchers in predicting the health of mechanical equipment. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] Therefore, the purpose of this invention is to propose a cyber-physical fusion method for predicting the health of mechanical equipment, which is beneficial for supporting predictive maintenance activities, thereby realizing the monitoring and control of mechanical equipment in network collaborative manufacturing and smart factories from physical components to network components and back to physical components in the cyber-physical fusion system (CPS).

[0008] Another aspect of the present invention is to propose a CPS hybrid modeling optimization system that employs a mechanical equipment health prediction method.

[0009] To achieve the above objectives, this invention proposes a cyber-physical fusion method for predicting the health of mechanical equipment, comprising:

[0010] Collect the working status data of the mechanical equipment; input the working status data into a pre-trained cyber-physical system (CPS) hybrid model to obtain the health prediction result of the mechanical equipment, wherein the CPS hybrid model is obtained by fusing physical modeling of signal processing and data modeling of deep learning; perform data mining on the health prediction result, and evaluate the real-time status of the mechanical equipment based on the mining results to obtain the status evaluation result of the machine equipment.

[0011] The mechanical equipment health prediction method of this invention introduces a time-frequency attention (TFA) mechanism in physical modeling, assigning different weights to different times and frequencies to acquire more discriminative and important features while ignoring redundant information. In data modeling, adaptive learning rate and label smoothing regularization methods are introduced to make model training more stable. The combination of physical and data modeling results in higher modeling accuracy and better prediction of mechanical equipment status. This facilitates the transition from preventative to predictive maintenance of mechanical equipment in network-based collaborative manufacturing and smart factories.

[0012] In addition, the mechanical equipment health prediction method according to the above embodiments of the present invention may also have the following additional technical features:

[0013] Furthermore, in one embodiment of the present invention, after collecting the working status data of the mechanical equipment, the method further includes: preprocessing the working status data, wherein the preprocessing includes one or more of data fusion processing, data cleaning processing, and data normalization processing.

[0014] Furthermore, in one embodiment of the present invention, before inputting the working status data into the pre-trained CPS hybrid model, the method further includes: acquiring a training dataset, wherein the training dataset includes various training sample data of mechanical equipment; extracting the time-frequency matrix of the training dataset; and, based on the weights of the time-frequency matrix, integrating the time-frequency matrices with different weights into a deep residual network ResNet to obtain a fused ResNet; and training the CPS hybrid model in the fused ResNet using the training dataset to obtain the trained CPS hybrid model.

[0015] Furthermore, in one embodiment of the present invention, the step of extracting the time-frequency matrix of the training dataset and, based on the weights of the time-frequency matrix, integrating the time-frequency matrices with different weights into a ResNet to obtain a fused ResNet includes: decomposing the time-frequency matrix to obtain aggregated features of time feature information and frequency feature information; performing channel transformation on the aggregated features to obtain time attention weights and frequency attention weights; and integrating the time attention weights and frequency attention weights into a ResNet to obtain the fused ResNet.

[0016] Furthermore, in one embodiment of the present invention, the step of training a CPS hybrid model in the fused ResNet using the training dataset to obtain the trained CPS hybrid model includes: applying a preset strategy to the learning rate of the fused ResNet to obtain the learning rate of the fused ResNet; and calculating the correct label position loss and the incorrect label position loss of the training dataset using the cross-entropy loss function based on the learning rate of the fused ResNet.

[0017] To achieve the above objectives, a second aspect of the present invention proposes a CPS hybrid modeling and optimization system employing a mechanical equipment health prediction method, comprising:

[0018] The perception layer is used to collect the working status data of the mechanical equipment; the hybrid modeling layer is used to input the working status data into a pre-trained cyber-physical system (CPS) hybrid model to obtain the health prediction result of the mechanical equipment, wherein the CPS hybrid model is obtained by fusing physical modeling of signal processing and data modeling of deep learning; the decision layer is used to perform data mining on the health prediction result, evaluate the real-time status of the mechanical equipment based on the mining results, and obtain the status evaluation result of the mechanical equipment.

[0019] This invention presents a CPS hybrid modeling and optimization system employing a mechanical equipment health prediction method. In physical modeling, the introduction of TFA (Transformative Facilitation) assigns different weights to different times and frequencies, thereby acquiring more discriminative and important features while ignoring redundant information. In data modeling, the introduction of adaptive learning rate and label smoothing regularization methods makes model training more stable. The combination of physical and data modeling results in higher modeling accuracy and better prediction of mechanical equipment status. This facilitates the transition from preventative to predictive maintenance of mechanical equipment in network-based collaborative manufacturing and smart factories.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0022] Figure 1 A flowchart illustrating a cyber-physical fusion method for predicting the health of mechanical equipment according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of an STFT according to an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of a TFA according to an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of the StepLR adaptive learning rate adjustment strategy according to an embodiment of the present invention;

[0026] Figure 5 This is a schematic diagram of the CosLR adaptive learning rate adjustment strategy according to an embodiment of the present invention;

[0027] Figures 6(a), 6(b), and 6(c) are respectively the ResNet structure diagram, TFA-ResNet structure diagram, and TFA-ResNet unfolded structure diagram according to an embodiment of the present invention;

[0028] Figure 7 This is a schematic diagram of the CPS hybrid modeling and optimization system structure using a mechanical equipment health prediction method according to an embodiment of the present invention;

[0029] Figure 8 This is a flowchart of the CPS (Continuous Health Prediction System) for mechanical equipment according to an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] The mechanical equipment health prediction method and the CPS hybrid modeling and optimization system using the mechanical equipment health prediction method are described below with reference to the accompanying drawings, according to embodiments of the present invention.

[0033] Example 1:

[0034] Figure 1 This is a flowchart of a mechanical equipment health prediction method according to an embodiment of the present invention.

[0035] like Figure 1 As shown, the method includes, but is not limited to, the following steps:

[0036] S1 collects the working status data of mechanical equipment.

[0037] S2, input the working status data into the pre-trained cyber-physical system (CPS) hybrid model to obtain the health prediction results of the mechanical equipment. The CPS hybrid model is obtained by fusing physical modeling of signal processing and data modeling of deep learning.

[0038] S3 performs data mining on the health prediction results, evaluates the real-time status of the machinery and equipment based on the mining results, and obtains the status assessment results of the machinery and equipment.

[0039] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] As an example, this invention introduces data feature extraction using the short-time Fourier transform (STFT). The principle of STFT is to multiply the original signal by a window function and then perform a piecewise Fourier transform on the original signal by moving the window function. Commonly used window functions for STFT include rectangular windows, Hamming windows, and Gaussian windows. The mathematical expression for STFT can be represented as:

[0041]

[0042] Where t and τ represent time, ω represents frequency, f(t) is the time-domain signal, h(τ-t) represents the window function, and STFT(t, ω) is the time-frequency matrix after STFT transformation. Assuming the input signal length is 1024, after passing through a Hamming window STFT of length 64, the output is a 33×33 time-frequency matrix. Figure 2 As shown, the horizontal axis represents time, the vertical axis represents frequency, and different colors represent different values.

[0043] Understandably, in practical engineering applications of mechanical equipment health prediction, due to the variable speed and operating conditions, the samples in the training sample set are often collected under many different speeds and operating conditions. Because of the significant differences in speed and operating conditions among the training samples, the differences in statistical characteristics of samples of the same class also increase, leading to relatively small differences between samples of different categories. This results in overlapping and indistinguishable samples of different types in high-dimensional space, severely affecting the classification performance of deep learning methods. In STFT-based physical modeling methods, under alternating speed and operating conditions, it is often difficult to predetermine which time points and frequency bands contain discriminative information and which contain a large amount of redundant information. This paper addresses the characteristics of the time-frequency matrix obtained by STFT.

[0044] This invention constructs a novel TFA to apply different weights to coefficients from different time points and frequency bands, thereby achieving adaptive adjustment of the importance of information from different time points and frequency bands, such as... Figure 3 As shown, the entire process consists of 4 steps.

[0045] Step 1: First, global average pooling is decomposed into average pooling operations along the time axis and frequency axis to capture temporal and frequency feature information, respectively. Given input X, X∈R T×F×c . Let represent the element in the i-th column and j-th row of the c-th channel of input X, where i ∈ [1, T] and j ∈ [1, F]. (T, F) represents the spatial dimension of input matrix X. The average value at different time points is obtained by performing average pooling on input matrix X along the frequency axis (vertical axis) and taking the average value of each column for time information encoding. The output of the c-th channel at time i can be expressed as:

[0046]

[0047] By performing average pooling on the input matrix X along the time axis (horizontal axis), the average value of each row is taken and used to encode the frequency information to obtain the average value of different frequency bands. The output of the c-th channel at frequency j can be expressed as:

[0048]

[0049] The two transformations described above aggregate features along both the temporal and frequency spatial directions, respectively. These two transformations allow our attention blocks to capture feature information along both the temporal and frequency directions, which helps the network to more accurately locate objects of interest.

[0050] Step 2: Concatenate the aggregated features from Step 1, and then feed them into a 1×1 convolution function Conv1 to obtain:

[0051] S=δ(Conv i ([z I , z J ]))

[0052] In the formula, [.,.] represents the splicing operation along the spatial dimension, δ represents the nonlinear activation function h swish, and S∈R c / r×(T+F) It is an intermediate feature map that encodes positional information in the time axis and frequency axis directions. r is the reduction ratio that controls the size of the S dimension, which helps to reduce the number of parameters.

[0053] Step 3: Split the S obtained in Step 2 into two independent tensors g∈R along the spatial dimension. c / r×T and h∈R c / r×F Then, using two other convolution operations, Conv... t and Conv fBy converting the channels of g and h to the same channel size as the input X, the temporal attention weight t can be obtained. c and frequency attention weight f c This can be expressed by an equation:

[0054] t c =a(Conv t (g))

[0055] f c =σ(Conv f (h))

[0056] In the formula, σ represents the sigmoid activation function.

[0057] Step 4: Apply the temporal attention weights t obtained in Step 3. c and frequency attention weight f c Multiplying the result by the input matrix X yields the output y after TFA processing. c (i, j) can be expressed by the equation as:

[0058]

[0059] Furthermore, the learning rate, as a crucial hyperparameter in deep neural networks, significantly impacts the performance of deep learning models. The learning rate controls the magnitude of each parameter update. If the learning rate is too large, the parameter update pace is too rapid, potentially exceeding the minimum loss value. Ultimately, the parameter values ​​may hover around their optimal values, making deep learning training difficult to converge. If the learning rate is too low, the corresponding parameter update pace is too slow, consuming more time resources to find the optimal value and easily getting trapped in local optima, also hindering model convergence. Therefore, exploring a suitable learning rate adjustment method for model training is necessary and valuable.

[0060] As an example, this invention employs two strategies for controlling the learning rate. Adjusting the learning rate at equal intervals of epochs is called StepLR, such as... Figure 4 As shown. Taking the cosine function as the period, and resetting the learning rate at the maximum value of each period, is called CosLR, as shown. Figure 5 As shown.

[0061] When using the cross-entropy loss function, only the loss at the correct label positions (positions with one-hot labels of 1) in the training samples is considered, while the loss at the incorrect label positions (positions with one-hot labels of 0) is ignored. This allows the model to fit the training set very well, but because the loss at other incorrect label positions is not calculated, the probability of prediction errors increases. To address this issue, label smoothing regularization (LSR) was developed. This helps to solve the problem of model overconfidence, reduces overfitting, and improves the model's learning ability.

[0062] Assuming p(k) is the predicted distribution and q(k) is the true distribution, and the true distribution after smoothing with coefficient ε and class K labels is q′(k), the relationship between the cross-entropy loss I0 and the label smoothing loss I can be expressed as:

[0063]

[0064] Learning smooth labels instead of true labels helps reduce overfitting, thereby improving model performance.

[0065] Finally, by embedding the obtained temporal attention weights and frequency attention weights into the deep learning network architecture ResNet, the CPS physical modeling process for predicting the health of mechanical equipment is optimized. Figure 6(a) , 6(b) As shown in Figure 6(c), Figure 6(a) is the traditional ResNet, Figure 6(b) is the ResNet with TFA added, and Figure 6(c) is the unfolded TFA-ResNet network structure. BN represents batch normalization, TFA represents the TFA module, ReLU represents the rectified linear activation function, and Conv represents the convolution operation. Then, during deep learning training, adaptive learning rate and label smoothing regularization are used to optimize the CPS data modeling process for predicting the health of mechanical equipment.

[0066] (1) This invention first proposes a novel attention mechanism for optimizing the physical modeling of CPS (Continuous Power Scheme) in predicting the health of mechanical equipment. This novel attention mechanism, called TFA, consists of a frequency attention mechanism and a time attention mechanism. Specifically, in traditional CPS hybrid modeling, physical modeling is mainly achieved through signal processing methods such as STFT (Signal-Thinning Tactics) and wavelet packet decomposition (WPD). TFA is used to weight the importance of information from different frequency bands and different time points, thereby enabling the extraction of more discriminative and important features from the data.

[0067] (2) Adaptive learning rate and label smoothing regularization methods were employed to optimize CPS data modeling for mechanical equipment health prediction. Specifically, in deep learning-based data modeling, methods such as equal-interval decay learning rate and cosine learning rate were used to train the model, thereby controlling the pace of model parameter updates and finding the optimal learning rate update method by comparing the results. Label smoothing regularization helps reduce model overfitting and optimizes the training process. These two methods improve the accuracy of CPS data modeling.

[0068] (3) A new CPS hybrid modeling method is proposed. By integrating TFA from (1) into the deep learning basic network architecture ResNet and adopting the training method proposed in (2), the CPS hybrid modeling is optimized and the modeling accuracy is improved.

[0069] (4) Further, the present invention collects real-time data of mechanical equipment by arranging various sensors (such as vibration sensors, current sensors, and noise sensors) at different locations in the workshop and production line; then, the collected industrial big data is preprocessed, including data fusion, data cleaning, and data normalization; then, the preprocessed data is sent into a hybrid model for further feature extraction and model training and prediction; finally, the meaningful prediction results are further analyzed and mined to evaluate the real-time status of the mechanical equipment. The evaluation results are conducive to supporting predictive maintenance activities, thereby realizing the health prediction of mechanical equipment CPS from physical components to network components and back to physical components in network collaborative manufacturing and smart factories.

[0070] In summary, this invention introduces a novel optimization method for CPS hybrid modeling of mechanical equipment health prediction in network collaborative manufacturing and smart factories. Currently, the main methods for CPS modeling include physical modeling, data modeling, and hybrid modeling. Pure physical modeling is only suitable for simple systems; for complex systems like mechanical equipment in network collaborative manufacturing and smart factories, it is difficult to establish accurate physical models. Pure data modeling, without any physical preprocessing, results in low modeling accuracy. Hybrid modeling without optimization also suffers from low accuracy. To address these issues, this invention proposes a method for optimizing CPS physical and data modeling of mechanical equipment. Specifically, in physical modeling, TFA is introduced to assign different weights to different times and frequencies, thereby acquiring more discriminative and important features and ignoring redundant information. In data modeling, adaptive learning rate and label smoothing regularization are introduced to make model training more stable. The combination of physical and data modeling results in higher modeling accuracy and better prediction of the mechanical equipment's condition. This helps transition the maintenance of mechanical equipment in network collaborative manufacturing and smart factories from preventative maintenance to predictive maintenance.

[0071] To verify the effectiveness of the method in this embodiment of the invention, data from three common key components in mechanical equipment—machine tool spindles, bearings, and gears—were selected as the dataset. Verification was conducted through machine tool spindle rotation error monitoring experiments and bearing and gear fault diagnosis experiments.

[0072] (1) Experimental Environment and Experimental Setup

[0073] The experiments were conducted on a system with an i7 CPU @ 3.80GHz and an NVIDIA GeForce RTX 2060 SUPER, using PyTorch as the programming language. Adam was used as the optimizer during model training. Momentum, an important parameter of Adam that accelerates the training process, was set to 0.9. The maximum number of epochs was set to 100. To verify the effectiveness of the proposed optimization method in CPS hybrid modeling, nine control experiments were set up for each dataset. FixLR represents a fixed learning rate of 0.001; StepLR represents adjusting the learning rate at equal intervals of epochs, multiplying the learning rate by 0.1 at epochs 30 and 60 respectively; CosLR represents a cosine learning rate with a period of 64; TFA represents physical modeling using a time-frequency attention mechanism; LSR represents data modeling using label smoothing regularization; and + represents a combination of centralized methods.

[0074] Method 1: ResNet+FixLR, a combination of ResNet and a fixed learning rate.

[0075] Method 2: ResNet+StepLR, a combination of ResNet and an equally spaced decaying learning rate, one of the data modeling optimization methods.

[0076] Method 3: ResNet+CosLR, a combination of ResNet and cosine learning rate, one of the data modeling optimization methods.

[0077] Method 4: ResNet+FixLR+TFA, a combination of ResNet, fixed learning rate and TFA physical modeling optimization method.

[0078] Method 5: ResNet+StepLR+TFA, a combination of ResNet, one of the data modeling optimization methods, with equal-interval decaying learning rate, and TFA physical modeling optimization method.

[0079] Method 6: ResNet+CosLR+TFA, a combination of ResNet, cosine learning rate (one of the data modeling optimization methods), and TFA (physical modeling optimization method).

[0080] Method 7: ResNet+FixLR+TFA+LSR, a combination of ResNet, fixed learning rate, TFA physical modeling optimization method, and label smoothing regularization method, one of the data modeling optimization methods.

[0081] Method 8: ResNet+StepLR+TFA+LSR, a combination of ResNet, TFA physical modeling optimization method and data modeling optimization method (equal interval decay learning rate and label smoothing regularization method).

[0082] Method 9: ResNet+CosLR+TFA+LSR, a combination of ResNet, TFA physical modeling optimization methods and data modeling optimization methods (cosine learning rate and label smoothing regularization method).

[0083] (2) Experimental comparison and result analysis

[0084] The experimental results on the machine tool spindle, bearing, and gear datasets are shown in Tables 1-3. The results show that the TFA-based physical modeling optimization method and the adaptive learning rate-based data modeling optimization method significantly improve the model accuracy. The label smoothing regularization method improves prediction accuracy on the spindle and bearing datasets, and achieves accuracy comparable to the cross-entropy loss function on the gear dataset. Therefore, it can be concluded that the TFA-based CPS physical modeling optimization method and the adaptive learning rate combined with label smoothing regularization-based data modeling optimization method can improve the accuracy of CPS hybrid modeling.

[0085] Table 1. Experimental results of machine tool spindle data condition monitoring.

[0086] LR FixLR StepLR CosLR FixLR StepLR CosLR FixLR StepLR CosLR TFA √ √ √ √ √ √ LSR √ √ √ accuracy 92.08 92.54 92.57 92.91 93.35 93.49 93.02 93.75 93.92

[0087] Table 2 Experimental results of bearing dataset fault diagnosis

[0088] LR FixLR StepLR CosLR FixLR StepLR CosLR FixLR StepLR CosLR TFA √ √ √ √ √ √ LSR √ √ √ accuracy 0.9386 0.9512 0.9485 0.9448 0.9552 0.9528 0.9466 0.9565 0.9528

[0089] Table 3. Experimental results of fault diagnosis for the gear dataset.

[0090] LR FixLR StepLR CosLR FixLR StepLR CosLR FixLR StepLR CosLR TFA √ √ √ √ √ √ LSR √ √ √ accuracy 0.9346 0.9398 0.9394 0.9431 0.9482 0.9478 0.9424 0.9480 0.9459

[0091] Example 2:

[0092] Secondly, a CPS hybrid modeling optimization system employing a mechanical equipment health prediction method according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0093] Figure 7 This is a schematic diagram of the structure of a CPS hybrid modeling optimization system using a mechanical equipment health prediction method according to an embodiment of the present invention. The system 10 includes: a perception layer 100, a hybrid modeling layer 200, and a decision layer 300.

[0094] Sensing layer 100 is used to collect working status data of mechanical equipment;

[0095] Hybrid modeling layer 200 is used to input working status data into a pre-trained cyber-physical system (CPS) hybrid model to obtain health prediction results for mechanical equipment. The CPS hybrid model is obtained by fusing physical modeling based on signal processing and data modeling based on deep learning.

[0096] The decision layer 300 is used to perform data mining on the health prediction results, and to evaluate the real-time status of the machinery and equipment based on the mining results, thereby obtaining the status evaluation results of the machinery and equipment.

[0097] Furthermore, following the aforementioned perception layer 100, a data preprocessing layer is also included.

[0098] This is used to preprocess work status data, where preprocessing includes one or more of the following: data fusion processing, data cleaning processing, and data normalization processing.

[0099] Furthermore, prior to the aforementioned hybrid modeling layer 200, the following is also included:

[0100] The sample acquisition module is used to acquire the training dataset, which includes various training sample data of mechanical equipment.

[0101] The network fusion module is used to extract the time-frequency matrix of the training dataset. Based on the weights of the time-frequency matrix, the time-frequency matrices with different weights are integrated into the deep residual network ResNet to obtain the fused ResNet.

[0102] The model training module is used to train the CPS hybrid model in the fused ResNet using the training dataset, so as to obtain the trained CPS hybrid model.

[0103] Furthermore, the aforementioned network convergence module includes:

[0104] The matrix decomposition module is used to decompose the time-frequency matrix to obtain aggregated features of time feature information and frequency feature information;

[0105] The feature transformation module is used to perform channel transformation on aggregated features to obtain temporal attention weights and frequency attention weights;

[0106] The weight fusion module is used to integrate temporal attention weights and frequency attention weights into ResNet to obtain a fused ResNet.

[0107] Furthermore, the aforementioned model training module includes:

[0108] The strategy preset module is used to execute a preset strategy on the learning rate of the fused ResNet to obtain the learning rate of the fused ResNet.

[0109] The loss calculation module is used to calculate the correct label location loss and incorrect label location loss of the training dataset using the cross-entropy loss function based on the learning rate of the fused ResNet.

[0110] Specifically, in practical applications, considering the characteristics of mechanical equipment in network collaborative manufacturing and smart factories, a CPS framework flowchart is constructed using a four-layer architecture, such as... Figure 8 As shown, it includes a perception layer 100, a data preprocessing layer, a hybrid modeling layer 300, and a decision layer 400. The physical space of CPS is embodied in the perception layer, while the information space is embodied in the data preprocessing layer, the hybrid modeling layer 300, and the decision layer 400.

[0111] Perception Layer 100: The key to industrial big data is how to collect data efficiently and reliably. This is achieved by deploying various sensors (such as vibration sensors, current sensors, and noise sensors) at different locations in workshops and production lines within networked collaborative manufacturing and smart factories to collect real-time data. Sensor deployment and data collection should be selective and focused based on specific needs.

[0112] Data Preprocessing Layer: The raw data collected by the perception layer 100 exhibits 3V characteristics. 3V stands for volume, variety, and velocity. Volume refers to the large amount of data that needs to be collected, processed, and stored; variety refers to the complexity of the data type, including structured and unstructured data. Structured data can be stored in tabular form, while unstructured data cannot. Velocity refers to the speed at which data is generated, requiring powerful data processing capabilities. Therefore, appropriate data preprocessing is essential. Data fusion can select key sensor data related to the state of mechanical equipment, which requires domain expert knowledge. Data cleaning can remove errors and redundant information from the data. Data normalization can eliminate the impact of inconsistent data distribution.

[0113] Hybrid Modeling Layer 200: This layer incorporates the aforementioned CPS physical modeling and CPS data modeling. Preprocessing based on physical relationships allows for the extraction of more valuable features from the original data. Data features derived from CPS physical modeling are more discriminative, which improves modeling accuracy. Adaptive learning rate and label smoothing regularization optimize the data modeling training process. The combination of these two methods makes CPS hybrid modeling for predicting the health of mechanical equipment more accurate.

[0114] Decision Layer 300: This layer mainly analyzes and mines the meaningful prediction results of Hybrid Modeling Layer 200, thereby assessing the real-time status of mechanical equipment. The assessment results are conducive to supporting predictive maintenance activities (anomaly detection, fault diagnosis), thereby realizing the feedback control of CPS for predicting the health of mechanical equipment from physical space to cyberspace and back to physical space.

[0115] The CPS hybrid modeling and optimization system of this invention, employing a mechanical equipment health prediction method, achieves higher accuracy in physical modeling by introducing TFA (Transformative Learning Aspect) to assign different weights to different times and frequencies, thereby acquiring more discriminative and important features and ignoring redundant information. In data modeling, the introduction of adaptive learning rate and label smoothing regularization methods makes model training more stable. The combination of physical and data modeling results in higher modeling accuracy and better prediction of mechanical equipment health. This facilitates the transition from preventative to predictive maintenance of mechanical equipment in network-based collaborative manufacturing and smart factories.

[0116] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0117] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0118] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An information-physical fusion method for mechanical equipment health prediction, characterized in that, The method comprises the following steps: Collecting working state data of a mechanical device; Inputting the working state data into a pre-trained information-physical system (CPS) hybrid model to obtain a health prediction result of the mechanical device, wherein the CPS hybrid model is obtained by fusing physical modeling of signal processing and data modeling of deep learning; Performing data mining on the health prediction result, and evaluating a real-time state of the mechanical device according to a mining result to obtain a state evaluation result of the mechanical device; After collecting the working state data of the mechanical device, the method further comprises: Preprocessing the working state data, wherein the preprocessing comprises one or more of data fusion processing, data cleaning processing and data normalization processing; Before inputting the working state data into the pre-trained CPS hybrid model, the method further comprises: Obtaining a training data set, wherein the training data set comprises a plurality of training sample data of the mechanical device; Extracting a time-frequency matrix of the training data set, fusing time-frequency matrices with different weights into a deep residual network (ResNet) based on weights of the time-frequency matrix, and obtaining a fused ResNet; Training a CPS hybrid model in the fused ResNet by using the training data set to obtain the trained CPS hybrid model; The extracting of the time-frequency matrix of the training data set, the fusing of the time-frequency matrices with different weights into the ResNet based on the weights of the time-frequency matrix, and the obtaining of the fused ResNet comprise: Decomposing the time-frequency matrix to obtain aggregated features of time characteristic information and frequency characteristic information; Performing channel conversion on the aggregated features to obtain time attention weights and frequency attention weights; Fusing the time attention weights and the frequency attention weights into the ResNet to obtain the fused ResNet; The training of the CPS hybrid model in the fused ResNet by using the training data set to obtain the trained CPS hybrid model comprises: Performing a preset strategy on a learning rate of the fused ResNet to obtain a learning rate of the fused ResNet; Based on the learning rate of the fused ResNet, using a cross-entropy loss function to calculate a loss of a correct label position and a loss of an error label position of the training data set.

2. A CPS hybrid modeling optimization system employing the mechanical equipment health prediction of claim 1, wherein, The system comprises: A perception layer configured to collect working state data of a mechanical device; A hybrid modeling layer configured to input the working state data into a pre-trained information-physical system (CPS) hybrid model to obtain a health prediction result of the mechanical device, wherein the CPS hybrid model is obtained by fusing physical modeling of signal processing and data modeling of deep learning; A decision layer configured to perform data mining on the health prediction result, and evaluate a real-time state of the mechanical device according to a mining result to obtain a state evaluation result of the mechanical device; After the perception layer, the system further comprises a data preprocessing layer, The working state data is preprocessed, wherein the preprocessing includes one or more of data fusion processing, data cleaning processing, and data normalization processing; Before the mixed modeling layer, further comprising: a sample acquisition module configured to acquire a training data set, wherein the training data set includes a plurality of training sample data of a mechanical device; a network fusion module configured to extract a time-frequency matrix of the training data set, fuse time-frequency matrices with different weights into a deep residual network (ResNet) based on weights of the time-frequency matrix, and obtain a fused ResNet; a model training module configured to train the CPS mixed model in the fused ResNet using the training data set, and obtain a trained CPS mixed model; The network fusion module comprises: a matrix decomposition module configured to decompose the time-frequency matrix to obtain aggregated features of time characteristic information and frequency characteristic information; a feature conversion module configured to convert the aggregated features into channel to obtain time attention weights and frequency attention weights; a weight fusion module configured to fuse the time attention weights and the frequency attention weights into the ResNet to obtain the fused ResNet; The model training module comprises: a strategy preset module configured to perform a preset strategy on a learning rate of the fused ResNet to obtain a learning rate of the fused ResNet; a loss calculation module configured to calculate correct label position loss and incorrect label position loss of the training data set using a cross-entropy loss function based on the learning rate of the fused ResNet.

Images